pub fn dequant_turbo8_block_to_bf16(
raw: &[u8],
bs: usize,
nkv: usize,
hd: usize,
out: &mut [bf16],
)Expand description
2026-09-25: A Turbo8 block (FP8 E4M3 data, BF16 group scales) to BF16.
Layout per block:
data: bs * nkv * hd bytes, one FP8 byte per element.
scales: bs * nkv * (hd / NVFP4_GROUP_SIZE) * 2 bytes, little-endian BF16.
The scale read matches nvfp4_dequant of
kernels/gb10/common/paged_decode_attn_turbo8.cu.